Restructure benchmarks skill and rename to kaggle-benchmarks

[nicholaskang-us]

This PR restructures the benchmarks skill to adhere to the Agent Skills standard and completes the truncated content.

What problem are we solving

• The benchmarks skill was previously nested under references, which is not an auto-activatable location for Agent Skills-compliant agents (added a new folder)
• Not properly formatted with the skills table with a description/version at the top
• There were no references pointing agents to the official SDK repository or DeepWiki documentation

What changes are proposed

• Renamed and relocated the skill directory to to match the folder-name standard for auto-activation
• Updated the YAML frontmatter with proper name (), rich description with keywords, and metadata fields
• Added links to official SDK resources so agents have a direct reference for writing benchmark task files

[dolaameng]

@nicholaskang-us I think @rosbo and @stevemessick has a plan to organize the skills. Let's consult them first.

[stevemessick]

How well does this play with kaggle auth login? (Do we really need both?)

[rosbo]

kaggle auth login authenticates you with Kaggle.

Then, using your Kaggle credentials, kaggle benchmarks auth fetches a model proxy token. You must be autenticated before you can fetch a model proxy token.

[nicholaskang-us]

i'm assuming this means this is fine and we don't need to make any change?

[rosbo]

This file is referenced from the main SKILL: https://github.com/Kaggle/kaggle-cli/blob/main/skills/SKILL.md.

This will cause a broken link.

The idea is to have one skill with references to the different resources (e.g. kernels, models, datasets, benchmarks, etc).

[nicholaskang-us]

to clarify, you mean that we just expect users to download the kaggle-cli skill ONLY? and they would say they want to write a benchmark task, which would invoke the main skill

[rosbo]

The current thinking followed the playwright-cli structure where you have a SKILL.md explaining cross-cutting CLI concerns (like auth/authz) and then link to references file stored in the references/ folder for each product. One explaining how to use benchmark, one for kernels, etc.

Following this structure: https://agentskills.io/home

But maybe it is best to really have separate skill per concrete task (like creating a benchmark)... We can experiment and see how the agent performs.

[rosbo]

I would suggest to put your skill focused on creating a benchmark task e2e in the kaggle-skills repo given it is cross-cutting (kaggle-cli & kaggle-benchmarks).